Rarity-Gated Context Conditioning for Offline Imitation Learning-Based Maritime Anomaly Detection

The paper introduces Rarity-Gated Feature-wise Linear Modulation (RGFiLM), a novel conditioning module designed to enhance contextual anomaly detection in maritime scenarios by addressing frequency bias in context distributions. RGFiLM utilizes a rarity score derived from empirical context distributions to modulate hidden features, improving decision stability and reducing false alarms in rare contexts. Evaluated on AIS motion sequences with environmental context, RGFiLM demonstrated superior performance in mean F1-False Positive Rate trade-offs compared to existing methods, highlighting its potential for practitioners focused on reliable anomaly detection in imbalanced datasets.